Possibilities for transfer of relevant data without revealing structural information

Summary

In this paper, we discuss how we safely exchanged proprietary data between third parties in the early years of predictive ADME/Tox model development. At that time, industry scientists wanted to evaluate predictive models, but were not willing to share their structures with software vendors. At the same time, model developers were willing to run the scientists’ structures through their models, but they were not willing to reveal which descriptors were important for a particular predictive model. We developed a process where scientists could perform calculations on a broad number of commercially available public descriptors and forward results as a property file, instead of their structures. Meanwhile, the model developer could extract descriptors used in the predictive model, run the model, and pass results back to the scientist. On the following pages, we discuss the pros and cons of this approach, and we address questions such as: Can structural information that is proprietary be compromised from descriptors in ADME/Tox models? Can ADME/Tox predictions be made purely from descriptors, without the explicit knowledge of chemical structures, proprietary or otherwise?